1/*
2 * reserved comment block
3 * DO NOT REMOVE OR ALTER!
4 */
5/*
6 * jfdctflt.c
7 *
8 * Copyright (C) 1994-1996, Thomas G. Lane.
9 * This file is part of the Independent JPEG Group's software.
10 * For conditions of distribution and use, see the accompanying README file.
11 *
12 * This file contains a floating-point implementation of the
13 * forward DCT (Discrete Cosine Transform).
14 *
15 * This implementation should be more accurate than either of the integer
16 * DCT implementations.  However, it may not give the same results on all
17 * machines because of differences in roundoff behavior.  Speed will depend
18 * on the hardware's floating point capacity.
19 *
20 * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
21 * on each column.  Direct algorithms are also available, but they are
22 * much more complex and seem not to be any faster when reduced to code.
23 *
24 * This implementation is based on Arai, Agui, and Nakajima's algorithm for
25 * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
26 * Japanese, but the algorithm is described in the Pennebaker & Mitchell
27 * JPEG textbook (see REFERENCES section in file README).  The following code
28 * is based directly on figure 4-8 in P&M.
29 * While an 8-point DCT cannot be done in less than 11 multiplies, it is
30 * possible to arrange the computation so that many of the multiplies are
31 * simple scalings of the final outputs.  These multiplies can then be
32 * folded into the multiplications or divisions by the JPEG quantization
33 * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
34 * to be done in the DCT itself.
35 * The primary disadvantage of this method is that with a fixed-point
36 * implementation, accuracy is lost due to imprecise representation of the
37 * scaled quantization values.  However, that problem does not arise if
38 * we use floating point arithmetic.
39 */
40
41#define JPEG_INTERNALS
42#include "jinclude.h"
43#include "jpeglib.h"
44#include "jdct.h"               /* Private declarations for DCT subsystem */
45
46#ifdef DCT_FLOAT_SUPPORTED
47
48
49/*
50 * This module is specialized to the case DCTSIZE = 8.
51 */
52
53#if DCTSIZE != 8
54  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
55#endif
56
57
58/*
59 * Perform the forward DCT on one block of samples.
60 */
61
62GLOBAL(void)
63jpeg_fdct_float (FAST_FLOAT * data)
64{
65  FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
66  FAST_FLOAT tmp10, tmp11, tmp12, tmp13;
67  FAST_FLOAT z1, z2, z3, z4, z5, z11, z13;
68  FAST_FLOAT *dataptr;
69  int ctr;
70
71  /* Pass 1: process rows. */
72
73  dataptr = data;
74  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
75    tmp0 = dataptr[0] + dataptr[7];
76    tmp7 = dataptr[0] - dataptr[7];
77    tmp1 = dataptr[1] + dataptr[6];
78    tmp6 = dataptr[1] - dataptr[6];
79    tmp2 = dataptr[2] + dataptr[5];
80    tmp5 = dataptr[2] - dataptr[5];
81    tmp3 = dataptr[3] + dataptr[4];
82    tmp4 = dataptr[3] - dataptr[4];
83
84    /* Even part */
85
86    tmp10 = tmp0 + tmp3;        /* phase 2 */
87    tmp13 = tmp0 - tmp3;
88    tmp11 = tmp1 + tmp2;
89    tmp12 = tmp1 - tmp2;
90
91    dataptr[0] = tmp10 + tmp11; /* phase 3 */
92    dataptr[4] = tmp10 - tmp11;
93
94    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
95    dataptr[2] = tmp13 + z1;    /* phase 5 */
96    dataptr[6] = tmp13 - z1;
97
98    /* Odd part */
99
100    tmp10 = tmp4 + tmp5;        /* phase 2 */
101    tmp11 = tmp5 + tmp6;
102    tmp12 = tmp6 + tmp7;
103
104    /* The rotator is modified from fig 4-8 to avoid extra negations. */
105    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
106    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
107    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
108    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
109
110    z11 = tmp7 + z3;            /* phase 5 */
111    z13 = tmp7 - z3;
112
113    dataptr[5] = z13 + z2;      /* phase 6 */
114    dataptr[3] = z13 - z2;
115    dataptr[1] = z11 + z4;
116    dataptr[7] = z11 - z4;
117
118    dataptr += DCTSIZE;         /* advance pointer to next row */
119  }
120
121  /* Pass 2: process columns. */
122
123  dataptr = data;
124  for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
125    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
126    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
127    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
128    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
129    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
130    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
131    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
132    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
133
134    /* Even part */
135
136    tmp10 = tmp0 + tmp3;        /* phase 2 */
137    tmp13 = tmp0 - tmp3;
138    tmp11 = tmp1 + tmp2;
139    tmp12 = tmp1 - tmp2;
140
141    dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
142    dataptr[DCTSIZE*4] = tmp10 - tmp11;
143
144    z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */
145    dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
146    dataptr[DCTSIZE*6] = tmp13 - z1;
147
148    /* Odd part */
149
150    tmp10 = tmp4 + tmp5;        /* phase 2 */
151    tmp11 = tmp5 + tmp6;
152    tmp12 = tmp6 + tmp7;
153
154    /* The rotator is modified from fig 4-8 to avoid extra negations. */
155    z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */
156    z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */
157    z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */
158    z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */
159
160    z11 = tmp7 + z3;            /* phase 5 */
161    z13 = tmp7 - z3;
162
163    dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
164    dataptr[DCTSIZE*3] = z13 - z2;
165    dataptr[DCTSIZE*1] = z11 + z4;
166    dataptr[DCTSIZE*7] = z11 - z4;
167
168    dataptr++;                  /* advance pointer to next column */
169  }
170}
171
172#endif /* DCT_FLOAT_SUPPORTED */
173